1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-vaddc/sse-mul32-ld32.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include <assert.h>
11
12 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/intrinsics-polyfill.h>
20 #include <xnnpack/vaddsub.h>
21
22
xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32(size_t n,const int8_t * input_a,const int8_t * input_b,int8_t * output,const union xnn_qs8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])23 void xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x32(
24 size_t n,
25 const int8_t* input_a,
26 const int8_t* input_b,
27 int8_t* output,
28 const union xnn_qs8_addsub_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 const __m128i va_multiplier = _mm_load_si128((const __m128i*) params->sse4_mul32.a_multiplier);
31 const __m128i vshift = _mm_loadu_si32(params->sse4_mul32.shift);
32 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4_mul32.output_zero_point);
33 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse4_mul32.output_min);
34 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse4_mul32.output_max);
35
36 __m128i vbias = _mm_cvtsi32_si128(params->sse4_mul32.b_multiplier[0] * (int32_t) *input_b);
37 vbias = _mm_shuffle_epi32(vbias, _MM_SHUFFLE(0, 0, 0, 0));
38 vbias = _mm_add_epi32(vbias, _mm_load_si128((const __m128i*) params->sse4_mul32.bias));
39 for (; n >= 32 * sizeof(int8_t); n -= 32 * sizeof(int8_t)) {
40 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
41 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
42 const __m128i va89AB = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 8));
43 const __m128i vaCDEF = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 12));
44 const __m128i vaGHIJ = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 16));
45 const __m128i vaKLMN = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 20));
46 const __m128i vaOPQR = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 24));
47 const __m128i vaSTUV = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 28));
48 input_a += 32;
49 input_b += 32;
50
51 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
52 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
53 __m128i vacc89AB = _mm_macc_epi32(va89AB, va_multiplier, vbias);
54 __m128i vaccCDEF = _mm_macc_epi32(vaCDEF, va_multiplier, vbias);
55 __m128i vaccGHIJ = _mm_macc_epi32(vaGHIJ, va_multiplier, vbias);
56 __m128i vaccKLMN = _mm_macc_epi32(vaKLMN, va_multiplier, vbias);
57 __m128i vaccOPQR = _mm_macc_epi32(vaOPQR, va_multiplier, vbias);
58 __m128i vaccSTUV = _mm_macc_epi32(vaSTUV, va_multiplier, vbias);
59
60 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
61 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
62 vacc89AB = _mm_sra_epi32(vacc89AB, vshift);
63 vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift);
64 vaccGHIJ = _mm_sra_epi32(vaccGHIJ, vshift);
65 vaccKLMN = _mm_sra_epi32(vaccKLMN, vshift);
66 vaccOPQR = _mm_sra_epi32(vaccOPQR, vshift);
67 vaccSTUV = _mm_sra_epi32(vaccSTUV, vshift);
68
69 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
70 const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point);
71 const __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(vaccGHIJ, vaccKLMN), voutput_zero_point);
72 const __m128i voutOPQRSTUV = _mm_adds_epi16(_mm_packs_epi32(vaccOPQR, vaccSTUV), voutput_zero_point);
73
74 __m128i vout0123456789ABCDEF = _mm_packs_epi16(vout01234567, vout89ABCDEF);
75 __m128i voutGHIJKLMNOPQRSTUV = _mm_packs_epi16(voutGHIJKLMN, voutOPQRSTUV);
76
77 vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min);
78 voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, voutput_min);
79
80 vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max);
81 voutGHIJKLMNOPQRSTUV = _mm_min_epi8(voutGHIJKLMNOPQRSTUV, voutput_max);
82
83 _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF);
84 _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV);
85 output += 32;
86 }
87 if XNN_UNLIKELY(n != 0) {
88 do {
89 const __m128i va0123 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a));
90 const __m128i va4567 = _mm_cvtepi8_epi32(_mm_loadu_si32(input_a + 4));
91 input_a += 8;
92
93 __m128i vacc0123 = _mm_macc_epi32(va0123, va_multiplier, vbias);
94 __m128i vacc4567 = _mm_macc_epi32(va4567, va_multiplier, vbias);
95
96 vacc0123 = _mm_sra_epi32(vacc0123, vshift);
97 vacc4567 = _mm_sra_epi32(vacc4567, vshift);
98
99 const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point);
100
101 __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567);
102 vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min);
103 vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max);
104
105 if XNN_LIKELY(n >= (8 * sizeof(int8_t))) {
106 _mm_storel_epi64((__m128i*) output, vout0123456701234567);
107 output += 8;
108 n -= 8 * sizeof(int8_t);
109 } else {
110 if (n & (4 * sizeof(int8_t))) {
111 *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567);
112 vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32);
113 output += 4;
114 }
115 if (n & (2 * sizeof(int8_t))) {
116 *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0);
117 vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16);
118 output += 2;
119 }
120 if (n & (1 * sizeof(int8_t))) {
121 *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0);
122 }
123 n = 0;
124 }
125 } while (n != 0);
126 }
127 }
128